How Open-RAIL Connects AI Models to Different Robots

A robot cannot simply wait each time its AI model needs to think. Vision-language-action (VLA) and world-action (WAM) models can produce actions in chunks, while physical control must continue between those outputs. China Mobile’s open-source Open-RAIL sits between the models and robot bodies to manage that gap—and to make it easier to use different models with different machines. 37

Keeping inference from interrupting control

Open-RAIL uses a server–client design. The server handles model inference, scheduling and data management; the robot-side client handles sensing and execution. Observation, inference and control run as separate asynchronous flows, allowing the robot to execute an action chunk while the model works on the next. That addresses the stop-start motion that can result when control waits for slower inference. 311

It then smooths motion at two points. A trajectory smoother reduces noise within each action chunk. A chunk fuser joins a newly returned chunk to the motion already under way, targeting abrupt changes at the boundary. These steps operate in the execution framework rather than requiring changes to the model’s weights. 211

Making models and robots interchangeable

On the hardware side, a lightweight abstraction layer standardizes how the system reads robot state and sends actions. Connecting another robot calls for a driver adapter rather than a rewrite of the higher-level execution logic. On the model side, a common inference interface and adapter routing reduce the code needed to connect another model. 37

The reported first-release scope is four robot backends and ten model variants. One published list names Agibot G1, Ti5 T170C, Navi WA2 and a LeRobot-based simulation backend, along with variants from the ACT, GR00T, RDT-1B, SmolVLA, AgiBot GO-1, π0 and TAO families. That list does not establish that every model works with every robot. 9

China Mobile reports that onboarding a new robot can fall from weeks to hours and that a new model may need roughly 50–100 lines of adapter code. It also reports a drop in joint-acceleration standard deviation from more than 10 rad/s² to about 0.1 rad/s². These are reported results, not independent benchmarks across the supported hardware. 711

Open-RAIL also folds execution-data collection into real robot runs so the resulting data can feed later model iterations. Its server–client split allows inference to be placed on the device, at the edge or in the cloud without changing robot-side application code. The evidence here does not independently establish the collection overhead or detail every recorded field. 5814

The Apache-2.0 code and documentation are linked from the Open-RAIL project page. The earlier VLA-RAIL preprint describes the asynchronous linker, trajectory smoother and chunk fuser. 92